What changed
Semiconductor research firm SemiAnalysis claims that the illicit export of over 1.5 gigawatts (GW) of Nvidia AI accelerators to China may have played a role in Jensen Huang's sudden departure from the company. The report estimates that a significant amount of restricted Nvidia hardware, including models like the H200 and B200, has reached China despite U.S. export controls. The firm notes that these exports are likely to have been concealed through methods such as forged documents and substitute servers.
What this means in practice
The scale of the alleged smuggling is significant, with estimates suggesting that between 1.5 million and 2.142 million GPUs could have been diverted to China. This volume of hardware is comparable to or exceeds the scale of xAI's Colossus operations. The report highlights that modern AI accelerators require substantial power and infrastructure, and the diversion of such hardware could provide Chinese developers with a competitive edge in AI development. SemiAnalysis warns that complete restrictions on Chinese access to American AI accelerators could lead to the adoption of domestic alternatives, undermining international collaboration and potentially weakening U.S. tech companies.
Limitations and availability
The report does not disclose the exact number of accelerators shipped to China or their split between product generations. It also does not provide direct evidence that U.S. intelligence agencies influenced Huang's departure or that he participated in unauthorized exports. The findings are based on estimates and case studies, with two notable examples involving individuals associated with Supermicro and other companies. The report emphasizes the need for a balanced approach to restrictions, allowing Chinese developers access to U.S. hardware under oversight while limiting the scale and capabilities of such access.
